On a quiet Tuesday morning, a single tweet from Menlo Ventures partner Deedy Das shifted the macro lens for anyone watching the intersection of AI and crypto. His estimate: five Chinese AI startups—Zhipu, DeepSeek, Kling, MiniMax, and Moonshot—generated a combined $2.6 billion in annual revenue. As a cross-border payment researcher who spends my days tracing liquidity cycles, this number isn't just a headline. It's a signal. It tells me that centralized compute is no longer a speculative narrative—it's a real asset class, consuming capital, talent, and user attention that once flowed toward decentralized alternatives.
Tracing the quiet resilience beneath the market means looking beyond the price charts of tokens like RNDR or AKT and asking: what does this $2.6B actually represent in the global liquidity map? The answer is a structural shift that most crypto native analysts are missing.
Context: The Global Liquidity Map Shifts
To understand the implications, I step back to the post-ETF landscape. Since Q1 2024, institutional capital has bifurcated into two main channels: Wall Street's Bitcoin ETF and the AI compute stack. The latter includes hyperscalers, data centers, and model providers. The $2.6B in revenue from these five Chinese companies is roughly equivalent to the entire market cap of a mid-tier crypto project. But unlike most crypto activity, this revenue is backed by actual utility—API calls, enterprise deployments, and subscription fees.
Based on my audit experience building cross-border payment rails for European banks, I've learned that liquidity doesn't just move—it follows trust. And trust in centralized AI is building faster than trust in decentralized compute. Between 2020 and 2022, I audited several DeFi protocols and saw firsthand how fragile trust can be when liquidity is shallow. The 2022 bridge preservation work taught me that resilience requires real economic activity, not just speculation.
These Chinese startups are demonstrating economic activity. DeepSeek's $500M revenue comes from offering API calls at a fraction of OpenAI's price. Zhipu's $1B relies on government contracts for sovereign AI. Kling's $500M is tied to video generation inside the Kuaishou ecosystem. Each of these represents a captive demand pool. For crypto's decentralized compute networks—Akash, Render, io.net—the question isn't whether they can match this scale, but whether they need to.
Core: The Centralized Compute Advantage
The core insight here is nuanced but critical: centralized AI compute is achieving economies of scale that decentralized alternatives cannot replicate due to coordination costs and latency requirements. During my 2018 post-bubble audit of the XRP Ledger, I learned that network latency and consensus delays are structural—they cannot be optimized away with faster hardware alone. Decentralized compute networks inherit similar overhead: node verification, dispute resolution, and token incentives add friction that centralized clusters avoid.
Let's break down the numbers. DeepSeek's $500M revenue, at their typical pricing of $0.14 per million tokens for the V2 model, implies approximately 3.6 quadrillion tokens processed annually. That's roughly 1.8 billion API calls per day. A decentralized network would need to coordinate thousands of nodes to handle that volume, each with variable uptime and performance. The latency alone would make real-time inference impractical. The unit economics of decentralized inference simply don't math out compared to subsidized centralized alternatives.
But here's the silent crisis beneath the surface: those centralized revenues are built on negative gross margins. DeepSeek's API pricing is widely understood to be unsustainable—it's a land-grab strategy, subsidized by venture capital. During my 2020 DeFi yield safety investigation, I saw the same pattern: protocols offering unsustainable yields to capture TVL, only to collapse when the subsidies ended. The difference? DeFi collapsed in months; AI compute can burn for longer because the underlying asset (compute) has intrinsic value beyond speculation.
The real infrastructure metric to watch is not revenue, but cash flow—specifically, whether these companies can reach positive unit economics before their funding runs dry. In 2022, I traced the quiet resilience of bridge protocols by monitoring their liquidity reserves rather than their TVL. The same approach applies here: track the ratio of inference cost to revenue for each company.
Contrarian: The Decoupling Thesis
The prevailing narrative in crypto is that decentralized compute will inevitably capture a large share of AI workloads because of censorship resistance, cost advantages, and data sovereignty. But the Chinese AI revenue data suggests the opposite: centralized compute is getting cheaper and faster, potentially decoupling AI demand from crypto's value proposition entirely.
Consider the decoupling thesis I'm now tracking. If centralized AI can deliver 10x lower cost per token than any decentralized alternative, then the primary value of blockchain for AI is not in compute supply, but in payment rails for autonomous agent economies. This aligns with my 2026 work integrating AI agents with blockchain for cross-border B2B transactions. We designed a micro-payment protocol that allowed agents to settle transactions in real-time—not because the compute needed to be decentralized, but because the settlement needed to be trustless and auditable.
The contrarian angle: Crypto's role in AI is not to compete on compute—it's to provide the accountability layer. The same way Visa doesn't compete with Amazon's data centers, crypto doesn't need to compete with OpenAI's clusters. Instead, blockchain should focus on the "human-in-the-loop" verification of AI actions, the transparent logging of model decisions, and the tokenization of data contributions. The $2.6B Chinese AI revenue validates that centralized compute works at scale; now crypto must pivot to what it does best: trust infrastructure.
Takeaway: Positioning for the Next Cycle
As I review my notes from the 2022 bear market bridge preservation work, I'm reminded that the most important positioning happens during chop. The current market is sideways, and these AI revenue numbers are a signal to reassess which crypto-AI projects have real product-market fit.
The quiet resilience we need to trace is not in AI token pumps, but in the infrastructure that ensures these centralized systems remain accountable. Projects building verifiable inference, decentralized data provenance, and audit trails for AI agents—these will be the ones that survive when the centralized subsidies dry up. The Chinese AI startups are building the engines; crypto should build the dashboards and safety checks.
The takeaway is not to fade AI tokens entirely, but to shift focus from compute supply to settlement and compliance rails. When the next crypto cycle arrives, the winners will be those who solved the trust problem, not the compute problem.
As I often sign off in my research: Quiet audits prevent loud collapses. Let's keep auditing where the real value flows.